Kastor: Turning Physics Foundation Models Into Efficient Generative PDE Simulators
qberthet · x · 2026-09-04
arXiv 2608.06107 (Google team) proposes Kastor, a comprehensive methodology for adapting a deterministic physics foundation model into an efficient, accurate generative surrogate:
- Two-stage inference: a large-stride causal auto-regressive model plus a non-causal temporal super-resolution network, cutting error accumulation while minimizing compute;
- Mean Prediction Regularization (MPR): a novel objective constraining the model to predict the deterministic distribution mean under null noise, dramatically improving FGN and diffusion-based emulators;
- Spatial gradient matching improves accuracy and physical fidelity as measured by power spectrum density.
The work addresses long-standing error accumulation and stochasticity issues in autoregressive ML emulators. The poster is teaching an "AI for scientific computing" course around this topic.
More from Research
- Vals AI launches SRE-Bench, a cybersecurity benchmark testing LLMs on binary reverse engineering — dyn___ · 2026-09-04
- Emergent Misalignment Is Predictable Generalization, Not a Magic "Evil Persona" — burny_tech · 2026-09-04
- RL-driven progress may hit a wall on out-of-distribution generalization, researcher argues — chris_j_paxton · 2026-09-04
- UCSD study: synthetic dialogues from non-conversational data bootstrap recommender systems — UCSanDiego · 2026-09-04
- New Lower Bounds Push Gradient Descent Acceleration Beyond the Nesterov Era — burny_tech · 2026-09-04
- Reef: Open-Source Infra for Self-Improving Agents Gains ~300 Stars in 2 Days — pliang279 · 2026-09-04